changelog: PR numbers on the Unreleased lines that lacked them

CHANGELOG.md

@@ -8,47 +8,47 @@ Each release's page on GitHub has its notes and measurements. See [the 1.0.0 rel
 
 - MOSS-Image's prompt encoder runs on Metal inside the `moss_image` family, in bf16 as published with fp32 sums where the reference sums, under the Metal kernel contract (`tf-kernel-contract --only "moss image"`).
 
-- GLM-5.3-Flash on CUDA with `--vision`: `TF_GLM_VISION_HOST=1` keeps the vision tower's bf16 weights in pinned host memory mapped into the device and copies each unit into one device slot just before it runs, so rank 0 keeps 240 MiB for the tower instead of 1.05 GiB and every rank gets that room back for KV pages and the prompt cache. Unset, nothing changes.
+- GLM-5.3-Flash on CUDA with `--vision`: `TF_GLM_VISION_HOST=1` keeps the vision tower's bf16 weights in pinned host memory mapped into the device and copies each unit into one device slot just before it runs, so rank 0 keeps 240 MiB for the tower instead of 1.05 GiB and every rank gets that room back for KV pages and the prompt cache (#1760). Unset, nothing changes.
 
-- A tool marked `"strict": true` is now held to its schema under `tool_choice: "auto"` (or none) on every family with a grammar runtime: the server compiles a grammar that triggers at the call markup, so a strict tool's arguments always validate. `TENSORFOLD_TOOL_GRAMMAR=1` holds every auto tools request the same way.
+- A tool marked `"strict": true` is now held to its schema under `tool_choice: "auto"` (or none) on every family with a grammar runtime: the server compiles a grammar that triggers at the call markup, so a strict tool's arguments always validate (#1688). `TENSORFOLD_TOOL_GRAMMAR=1` holds every auto tools request the same way.
 
-- Qwen3.5 and 3.6 on CUDA serve past their trained 262,144 positions with static YaRN rope: `--rope-yarn 4` on the checkpoint as downloaded, or a config whose rope carries a YaRN block (SGLang's staged 1M copy loads as is), raise the window to 1,048,576. On one GB10, `nvidia/Qwen3.6-35B-A3B-NVFP4` prefills a 1,048,512-token prompt in 26 minutes, finds a passphrase placed 5% in, and decodes to the window's last token; drafted, concurrent and resumed replies equal their plain runs. Without YaRN every reply is byte-identical to before. HIP, Metal, the DFlash2 drafters and every other family refuse a scaled rope by name instead of reading it as plain rope.
+- Qwen3.5 and 3.6 on CUDA serve past their trained 262,144 positions with static YaRN rope: `--rope-yarn 4` on the checkpoint as downloaded, or a config whose rope carries a YaRN block (SGLang's staged 1M copy loads as is), raise the window to 1,048,576. On one GB10, `nvidia/Qwen3.6-35B-A3B-NVFP4` prefills a 1,048,512-token prompt in 26 minutes, finds a passphrase placed 5% in, and decodes to the window's last token; drafted, concurrent and resumed replies equal their plain runs. Without YaRN every reply is byte-identical to before. HIP, Metal, the DFlash2 drafters and every other family refuse a scaled rope by name instead of reading it as plain rope (#1675).
 
 - DeepSeek-V4.x requests that omit sampling fields now sample at the release's recommended settings, temperature 1.0 and top_p 1.0 from DeepSeek's model cards, instead of decoding greedily when the checkpoint has no `generation_config.json`. A `generation_config.json`, the serve flags and the request still override them (#1725).
 
-- DeepSeek-V4.x: `tool_choice: "required"` and named tool calls are now constrained to the model's own DSML call markup, so those requests are held to a valid call on DeepSeek too.
+- DeepSeek-V4.x: `tool_choice: "required"` and named tool calls are now constrained to the model's own DSML call markup, so those requests are held to a valid call on DeepSeek too (#1684).
 
-- Qwen3.6-35B-A3B on CUDA: `--drafter` also takes a DFlash2 draft model, chosen by the `dflash_config` in its `config.json`; the checkpoint's own directory still drafts with its MTP layer. With `incoai/Qwen3.6-35B-A3B-DFlash2` over `nvidia/Qwen3.6-35B-A3B-NVFP4`, one GB10 decodes 130 tok/s at one stream on a 4.2K-token prompt (MTP: 99), and every drafted reply equals its plain run.
+- Qwen3.6-35B-A3B on CUDA: `--drafter` also takes a DFlash2 draft model, chosen by the `dflash_config` in its `config.json`; the checkpoint's own directory still drafts with its MTP layer. With `incoai/Qwen3.6-35B-A3B-DFlash2` over `nvidia/Qwen3.6-35B-A3B-NVFP4`, one GB10 decodes 130 tok/s at one stream on a 4.2K-token prompt (MTP: 99), and every drafted reply equals its plain run (#1674).
 
 
 
-- A fresh-start session disk now clears the `.tfs` and `.tmp` files earlier serving processes left behind, under an exclusive lock on its directory, so restarts no longer fill the disk. Ported by @jayleaton from @mikolaj92's downstream fix, corroborated by @Jenpo.
+- A fresh-start session disk now clears the `.tfs` and `.tmp` files earlier serving processes left behind, under an exclusive lock on its directory, so restarts no longer fill the disk (#1680). Ported by @jayleaton from @mikolaj92's downstream fix, corroborated by @Jenpo.
 
-- GLM-5.3-Flash on CUDA: `TF_GLM_KV_SHARE=n` with `TF_KV_PAGES=on` keeps prompts as the KV pool's pages in the shared prompt store (the one `TF_QWEN_KV_SHARE` uses), with at most n states past them, on one GPU and at tensor parallel. Where main's flat store has no room for a long state (GB10s at TP=3 with the 1M window), a 64k message sent again resumes in about 1.7 s instead of prefilling for 36 s; token-id prompts resume at every width (#1510). Unset, nothing changes.
+- GLM-5.3-Flash on CUDA: `TF_GLM_KV_SHARE=n` with `TF_KV_PAGES=on` keeps prompts as the KV pool's pages in the shared prompt store (the one `TF_QWEN_KV_SHARE` uses), with at most n states past them, on one GPU and at tensor parallel. Where main's flat store has no room for a long state (GB10s at TP=3 with the 1M window), a 64k message sent again resumes in about 1.7 s instead of prefilling for 36 s; token-id prompts resume at every width (#1732, #1510). Unset, nothing changes.
 
 - MOSS-Image-Base's transformer runs on CUDA too, through the same test tool (`tf-moss-image`) and cases as on Metal (#1528, for #1518). It runs on an RTX 5090 (sm_120) and a GB10 (sm_121) and refuses other GPUs; the prompt encoder and image decoder stay in torch, and nothing is served yet. Against the publisher's torch bf16 pipeline, every step's cosine is 0.99967 or better, the same step run twice gives the same bytes, and a 1024x1024 picture is 44.0 dB (RTX 5090) and 45.9 dB (GB10) PSNR from torch's. Its 50 guided steps at 1024x1024 take 24.6 s on an RTX 5090 and 69.6 s on a GB10, 1.26x and 2.03x faster than torch. Its four kernels and full attention at head dim 128, a new instance of the vision towers' kernel, pass the kernel contract with five planted controls red; the towers' existing attention kernels compile to the same SASS. See [the recipe](docs/recipes/moss-image-cuda.md). Thanks to @BobClawblaw.
 
-- DeepSeek-V4.x replies carry the Spark recipe's per-response stats (`prompt`, `cached`, `ttft_s`, `finish`, `completion`, `prefill_s`, `decode_s`, `rounds`, `drafted`, `accepted`) in their `tensorfold` block, after its own keys; `TENSORFOLD_WIRE=spark` turns them on for any family and `native` turns them off.
+- DeepSeek-V4.x replies carry the Spark recipe's per-response stats (`prompt`, `cached`, `ttft_s`, `finish`, `completion`, `prefill_s`, `decode_s`, `rounds`, `drafted`, `accepted`) in their `tensorfold` block, after its own keys; `TENSORFOLD_WIRE=spark` turns them on for any family and `native` turns them off (#1753).
 
-- DeepSeek-V4.1 on CUDA backs up only the ratio-2 recurrent carries before each window, instead of every layer's sliding-window ring (89 device copies, 6.4 MB a window on Flash-1017); rejected rows still roll back exactly.
+- DeepSeek-V4.1 on CUDA backs up only the ratio-2 recurrent carries before each window, instead of every layer's sliding-window ring (89 device copies, 6.4 MB a window on Flash-1017); rejected rows still roll back exactly (#1752).
 
-- DeepSeek-V4.1's DSpark drafting on CUDA no longer drains the whole context before each draft pass's staging upload, so a draft pass stops waiting for every side stream.
+- DeepSeek-V4.1's DSpark drafting on CUDA no longer drains the whole context before each draft pass's staging upload, so a draft pass stops waiting for every side stream (#1751).
 
-- DeepSeek-V4.1 at tensor parallel: every rank's session store now charges disk eviction the same logical size, so the ranks evict the same prefixes and their caps no longer drift apart, and a restore checks the leader's identity before it changes anything. Adapted by @jayleaton from @florianbrede-ayet's downstream fix, with @rootsbymenda's report of the diverging caps.
+- DeepSeek-V4.1 at tensor parallel: every rank's session store now charges disk eviction the same logical size, so the ranks evict the same prefixes and their caps no longer drift apart, and a restore checks the leader's identity before it changes anything (#1679). Adapted by @jayleaton from @florianbrede-ayet's downstream fix, with @rootsbymenda's report of the diverging caps.
 
-- GLM-5.3-Flash on CUDA: `TF_GLM_HOST_CACHE=1` on every rank keeps its prompt states in pinned host memory mapped into the device, sized by `--prompt-cache-gib`, so on discrete cards where the KV pool leaves the device arena a fraction of a GiB a long conversation's state is kept instead of refused and the next turn resumes rather than prefilling the whole conversation (#1735, #1510). Unset, nothing changes.
+- GLM-5.3-Flash on CUDA: `TF_GLM_HOST_CACHE=1` on every rank keeps its prompt states in pinned host memory mapped into the device, sized by `--prompt-cache-gib`, so on discrete cards where the KV pool leaves the device arena a fraction of a GiB a long conversation's state is kept instead of refused and the next turn resumes rather than prefilling the whole conversation (#1736, #1735, #1510). Unset, nothing changes.
 
 - EXL3 mixture-of-experts models on CUDA (GLM-5.3-Flash and the Qwen3.5-MoE EXL3 packs): on sm_120 and sm_121 (RTX PRO and RTX 50 Blackwell, GB10), the grouped experts kernel reads a warp's next tile of weights while it decodes the current one (#1505). The decode and the sums are unchanged, so outputs are the same bits; the kernel contract checks them against the previous loop at one to three tiles a warp and at the served models' widths, with two planted load faults that must fail. On a GB10 the kernel ran 11% faster at MiMo-V2.6-Flash's expert shapes (109 to 97 µs a launch), and GLM-5.3-Flash at two GB10s decoded the same tokens, within 1% of its previous speed. Ampere and Ada GPUs build the same code as before. Thanks to @ajensenwaud.
 
-- CUDA graphs: a capture whose end fails because it was invalidated (`CUDA_ERROR_STREAM_CAPTURE_INVALIDATED`) no longer leaves its stream marked as capturing, which had refused every later capture on that stream and every temporary host upload in the process with `UnownedCaptureSource`. A failed end that leaves the stream capturing, such as one from the wrong thread, still keeps the guard. Thanks to @jayleaton.
+- CUDA graphs: a capture whose end fails because it was invalidated (`CUDA_ERROR_STREAM_CAPTURE_INVALIDATED`) no longer leaves its stream marked as capturing, which had refused every later capture on that stream and every temporary host upload in the process with `UnownedCaptureSource`. A failed end that leaves the stream capturing, such as one from the wrong thread, still keeps the guard (#1744). Thanks to @jayleaton.
 
 - Chat streams with thinking on: the joined `delta.reasoning_content` pieces now equal the finished reply's `reasoning_content` byte for byte. The stream had kept the whitespace the finished reply strips, such as the newline before `</think>`; it now holds whitespace back until more reasoning follows. A streamed `/v1/responses` reasoning item now matches the non-streamed one too. Thanks to @jayleaton.
 
-- Spark 2.5 on CUDA keeps its full-attention layers' K/V (every fourth layer) in pages of the core pool, leased per stream as it grows, while the sliding layers keep their rings; streams no longer reserve a full context each up front. On a GB10 tokens are unchanged and speed is level. `TENSORFOLD_SPARK25_KV_PAGES=0` keeps the linear caches.
+- Spark 2.5 on CUDA keeps its full-attention layers' K/V (every fourth layer) in pages of the core pool, leased per stream as it grows, while the sliding layers keep their rings; streams no longer reserve a full context each up front. On a GB10 tokens are unchanged and speed is level. `TENSORFOLD_SPARK25_KV_PAGES=0` keeps the linear caches (#1653).
 
 - Qwen3.8-27B on M5 Ultra: a prompt chunk of more than 128 rows runs each 4-bit projection in one dispatch instead of 128-row pieces, so each weight matrix is read once per chunk. Replies are token-identical; decode, verify rows and other chips are unchanged (#1738). Thanks to @Nipale-ai.
 
-- The server reads a request's top-level `thinking` switch (`{"type": "disabled"}` or `"enabled"`), as DeepSeek's own API and Hermes Agent's DeepSeek profile send it. Before, it was ignored and the prompt still opened a think block. `chat_template_kwargs` still wins when both are set.
+- The server reads a request's top-level `thinking` switch (`{"type": "disabled"}` or `"enabled"`), as DeepSeek's own API and Hermes Agent's DeepSeek profile send it. Before, it was ignored and the prompt still opened a think block. `chat_template_kwargs` still wins when both are set (#1742).
 
 - MiniMax-M3.1-Flash on CUDA, at tensor parallel over two GB10s (sm_121 and two ranks only; anything else is refused), from Mia-AiLab's published EXL3 pack read as stored: a lossy third-party quantization whose logits against the release, teacher-forced over 2,406 tokens, give a mean KL of 0.0788 and 89.9% top-1 agreement (#1388, after #1235). It serves the release's DSpark drafts, kept prompt states, tools, structured output and images; one stream with drafts ran 24 to 39 tokens a second. The shared EXL3 grouped experts gain a SwiGLU-OAI epilogue (`act_mode` 2); modes 0 and 1 are bit-equal to before in the kernel contract. Thanks to @BobClawblaw.
 
@@ -60,7 +60,7 @@ Each release's page on GitHub has its notes and measurements. See [the 1.0.0 rel
 
 - Flash Next on CUDA: `TF_FLASHNEXT_CHUNK_SOLO` (opt-in, up to 16,384 rows) lets a prompt pass take a larger chunk when no other stream is decoding, and `TF_FLASHNEXT_CHUNK`'s rows otherwise. On a GB10 at 4,096 rows, a lone prompt's first token comes 5-8% sooner, for 2.25 GiB more scratch and up to about 2 s of wait for a request that arrives during a large pass. Unset, nothing changes (#1733, option C of #1554). Thanks to @jschmied.
 
-- Qwen3.8-27B on Metal: the DFlash2 drafter reads its sliding-window keys and values in place in its cache while the context has not wrapped, instead of copying them to scratch for every draft layer each round. Drafts are bit-identical. Thanks to @kieranscotney.
+- Qwen3.8-27B on Metal: the DFlash2 drafter reads its sliding-window keys and values in place in its cache while the context has not wrapped, instead of copying them to scratch for every draft layer each round. Drafts are bit-identical (#1727). Thanks to @kieranscotney.
 
 - Flash Next's MTP head norms its 10,240-value rows with tiled launches on CUDA sm_120 and sm_121 and split wave launches on HIP gfx1201, storing the same bytes as before: the norm kernel runs 1.4x to 2.5x faster on a GB10 and 4x to 5x on a Radeon AI PRO R9700 at 1 to 16 rows (#1702, fixes #1689). Thanks to @laudney.